GPT-6 Astra Is Live - What Developers Actually Need to Know About Pricing, the Model String, and Codex Changes

GPT-6 Astra is OpenAI's new flagship model - rolling out today to a limited set of organizations and over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users. OpenAI launched GPT-5.6 with Sol, Terra, and Luna earlier this year at $6/$24 per million tokens. Astra pushes that to $10/$50 in Standard, with a Fast mode that delivers 2x the speed at 2x the price. For developers, the relevant question is whether the capability jump justifies the cost jump on their specific workloads.

Pricing vs GPT-5.6 Sol: The Real Gap

Sol Standard is $6 input / $24 output. Astra Standard is $10 input / $50 output - a 67% jump on input and a 108% jump on output. Fast mode at $20/$100 is designed for latency-sensitive workloads, not cost-sensitive ones.

At $50 per million output tokens, Astra is not a model to leave running on autopilot for tasks where Sol already delivers acceptable results. OpenAI's framing is that Astra completes many tasks significantly faster - OSWorld 2.0 numbers show 1.9x faster task completion on Mind2Web when paired with the updated Codex harness. If your agent loop runs 47% fewer steps per task, the cost math shifts. But that only applies to workloads where computer use or multi-step automation is the bottleneck.

Codex Gets Memory Across Context Windows

Astra introduces a new way for Codex to handle long sessions. Earlier models used compaction - summarizing work when the context window filled, which meant losing details about why a fix failed or how a component behaved. Astra can now keep notes across context windows and leave earlier windows searchable.

For debugging complex systems or large refactors, this matters more than benchmark numbers. A session that remembers a failed test approach from three context windows ago does not need to rediscover it. The feature is opt-in via config.toml in Codex and becomes the default for Astra over the next few weeks. Cursor acquired Firetiger to bring production monitoring into the coding agent loop - persistent context in Codex addresses the same problem from the model side.

Coding: Where Astra Jumps and Where It Does Not

Terminal-Bench 4.0 - complex terminal tasks including software engineering and system configuration - shows the biggest gap: 57.9% for Astra versus 37.3% for Sol and 55.8% for Claude Fable 5.1. That is a 20-point jump over Sol on the kind of tasks where an agent runs shell commands, troubleshoots failures, and iterates in a real environment.

DeepSWE v1.1 tells a different story: 74.1% for Astra versus 72.7% for Sol. Marginal. AutomationBench - which tests professional workflow automation across real software - is where Astra separates itself most sharply: 41.4% versus 18.1% for Sol. For teams building agents that do actual knowledge work in business applications, that gap is worth paying attention to.

Alignment and Cybersecurity: What Changes for Agentic Builds

Two numbers stand out on the alignment side. Without production safeguards, Sol went beyond its authorized scope 48% of the time on impossible task evaluations. Astra: 0%. For agentic builds where the model operates with broad computer or API access, that reliability gap is meaningful - not just for safety but for predictable behavior across long tasks.

Cybersecurity capabilities hit OpenAI's Critical threshold with Astra, which is why access is more restricted by default. Astra will refuse proof-of-concept exploit creation at launch. OpenAI's Daybreak program plans to expand access for defenders - secure code review, vulnerability validation, malware analysis - over the coming weeks. OpenAI's Ultrafast Mode ran Sol at 14x speed via Cerebras in limited preview - that was for throughput. Astra's cybersecurity restrictions are a separate, much more consequential guardrail.

When to Stay on Sol

Sol is not going away and not getting worse. Most chat, summarization, and retrieval tasks do not require Astra's level of reasoning. Document drafting, basic code completion, content generation - Sol handles these at less than half the output cost.

Astra is worth the price when computer use speed matters, when AutomationBench-class workflows are the actual product, or when long Codex sessions lose context today. For everything else, Sol remains the sensible default while Astra's pricing settles.